Breaking

TypeSafe's Jev Becomes a Control Layer for Agents, Up to 193.6× Faster in Company Tests

September 21, 2026 at 12:12 EDT

Foundation Models
AI Agents

TypeSafe AI is positioning Jev, its new decision model, as a control layer inside software agents, and it reports those decisions running up to 193.6 times faster and 444.6 times cheaper than the language models in its own workflow tests. The pattern now called Jev Engineering leaves drafting to a generative model, hands bounded forks to Jev, and lets ordinary code enforce the rules.

21 Sep 2026 · TypeSafe AI

A control layer that does not write

Jev answers choices, scores, and yes-or-no questions in one parallel pass, so agent code can route and approve without a chatbot in every loop. Company workflow tests put the top gap at 193.6× on time and 444.6× on cost.

193.6×

Top speed gap in the workflow tests

444.6×

Top cost gap in those same tests

0.4s

Jev time per workflow case

$0.042

Per million input tokens. Output is free.

Seconds per case, same four workflows

The tall column is the slow model

Bar height equals seconds. Jev at 0.4s is about one pixel beside Claude Sonnet 5 at 78.1s, near the 193.6× top of the test.

0.4s

10.1s

37.8s

78.1s

Jev
1×

GPT-5.6 Terra
~25×

Claude Opus 5
~95×

Claude Sonnet 5
~195×

Takeaway: the homepage speed claim is this high end, Sonnet 5's 78.1 seconds against Jev's 0.4, not a single cheap classification.

Dollars per case, separate scale

Cost, drawn the same way

Bar height equals dollars. Jev at $0.0004 is a hairline beside Claude Opus 5 at $0.1761, near the 444.6× top.

$0.0004

$0.0304

$0.1174

$0.1761

Jev
1×

Terra
~76×

Sonnet 5
~294×

Opus 5
~440×

How a loop is split

Write, decide, then act

1 · StateGoal, evidence, and what is still missing

→

2 · JevChoice, score, or yes-or-no, in parallel

→

3 · GateCode keeps only answers above its threshold

→

4 · ActA model writes, or a tool runs, if code allows it

Where builders say it fits

Repeated forks: route a model, score a source, classify a queue, or check a tool call. Reported runs include 1,018 papers for $0.08 and a flight search in about 7 seconds for $0.0039.

Where the multiple shrinks

One yes-or-no against a small model was about 1.7×. Invoice agreement was 61.8% against 79.1% for GPT-5.6 Sol. A typed answer can still be the wrong valid option.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.